[PATCH 10/13] x86/tdx: Convert shared memory back to private on kexec

[Kirill A. Shutemov]

On Fri, Oct 06, 2023 at 07:58:03AM -0700, Sean Christopherson wrote:
> On Thu, Oct 05, 2023, Kirill A. Shutemov wrote:
> > diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
> > index 7368d254d01f..b5acf9fb4c70 100644
> > --- a/arch/x86/Kconfig
> > +++ b/arch/x86/Kconfig
> > @@ -884,6 +884,7 @@ config INTEL_TDX_GUEST
> >  	select X86_MEM_ENCRYPT
> >  	select X86_MCE
> >  	select UNACCEPTED_MEMORY
> > +	select EMERGENCY_VIRT_CALLBACK
> >  	help
> >  	  Support running as a guest under Intel TDX.  Without this support,
> >  	  the guest kernel can not boot or run under TDX.
> 
> ...
> 
> >  void __init tdx_early_init(void)
> >  {
> >  	struct tdx_module_args args = {
> > @@ -882,6 +1007,14 @@ void __init tdx_early_init(void)
> >  	 */
> >  	x86_cpuinit.parallel_bringup = false;
> >  
> > +	machine_ops.shutdown = tdx_shutdown;
> > +
> > +	/*
> > +	 * KVM overrides machine_ops.crash_shutdown, use emergency
> 
> This is going to be super confusing.  KVM utilizes the emergency virt callback.
> The KVM paravirt guest code uses .crash_shutdown().  People that are passingly
> familiar with virt and know what KVM is, but don't already know the difference
> between the two are going to be all kinds of confused.
> 
> I also feel like you're playing with fire, e.g. what's to stop the hypervisor
> specific paravirt guest support from using .shutdown() in the future?
> 
> And the callback is invoked for far more than just kexec().  I don't see how the
> host can emulate a reboot without destroying and rebuilding the VM, e.g. it can't
> stuff register state to emulate INIT or RESET.  Unless I'm missing something,
> converting shared memory back to private for a shutdown or reboot is undesirable
> as adds one more thing that can go wrong and prevent the system from cleanly
> shutting down ASAP (for some definitions of "cleanly").

Okay, fair enough. I will look for better way to hookup into kexec
process. That was the best fit I found so far, but yes it is not ideal.

> Lastly, doesn't SEV need similar behavior?  This seems like core functionality
> for any guest with cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT).  Why not make the
> "unshare on kexec" code common and gate it with CC_ATTR_GUEST_MEM_ENCRYPT?

I don't know SEV specifics. I am open to collaboration on this.

Tom, Ashish, let me know if you need this in generic code. I can arrange
that.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov